# Load libraries
library(magrittr)
library(readr)
library(dplyr)
library(tidyr)
library(readxl)
library(stringr)
library(ggplot2)
library(gridExtra)
library(scales)
library(Cairo)
library(grid)
library(vcd)
library(countrycode)
library(maptools)
library(rgdal)
# Load data and add labels
responses.orgs <- readRDS(file.path(PROJHOME, "data_raw",
"responses_orgs_clean.rds"))
responses.countries <- readRDS(file.path(PROJHOME, "data_raw",
"responses_countries_clean.rds"))
responses.orgs.labs <- read_csv(file.path(PROJHOME, "data_raw",
"response_orgs_labels.csv"))
responses.countries.labs <- read_csv(file.path(PROJHOME, "data_raw",
"response_countries_labels.csv"))
Hmisc::label(responses.orgs, self=FALSE) <- responses.orgs.labs$varlabel
Hmisc::label(responses.countries, self=FALSE) <- responses.countries.labs$varlabel
# Add survey sources
phone <- readRDS(file.path(PROJHOME, "data_raw", "phone.rds"))
linkedin <- readRDS(file.path(PROJHOME, "data_raw", "linkedin.rds"))
responses.orgs %<>%
mutate(survey.method = ifelse(survey.id %in% phone, "Phone",
ifelse(survey.id %in% linkedin, "LinkedIn",
"Online")))
# Add a nicer short ID to each response to reference in the article
set.seed(1234)
nicer.ids <- data_frame(survey.id = responses.orgs$survey.id) %>%
mutate(clean.id = 1000 + sample(1:n()))
write_csv(nicer.ids, path=file.path(PROJHOME, "data", "id_lookup_WILL_BE_OVERWRITTEN.csv"))
The data is split into two sets: responses.orgs has a row for each surveyed organization and responses.countries has a row for each country organizations responded for (1-4 countries per organization). For ease of analysis, this combines them into one larger dataframe (so organization-level data is repeated). It also removes columns that were added manually, where an RA coded whether a country was mentioned in different questions (with a colum for each country!).
responses.all <- responses.orgs %>%
left_join(responses.countries, by="survey.id") %>%
select(-contains("_c", ignore.case=FALSE)) %>% # Get rid of all the dummy vars
left_join(nicer.ids, by="survey.id") %>%
select(survey.id, clean.id, everything())
Convert some responses into numeric indexes:
importance <- data_frame(Q3.19 = levels(responses.countries$Q3.19),
importance = c(2, 1, 0, NA))
positivity <- data_frame(Q3.25 = levels(responses.countries$Q3.25),
positivity = c(-1, 1, 0, NA))
improvement <- data_frame(Q3.26 = levels(responses.countries$Q3.26),
improvement = c(1, 0, -1, NA))
# Cho data
# TODO: Someday get this directly from the internet, like Freedom House data
# * http://www.economics-human-trafficking.org/data-and-reports.html
# Except that data doesn't include tier scores, so that would need to come
# from somewhere else...
tip.change <- read_csv(file.path(PROJHOME, "data", "policy_index.csv")) %>%
group_by(countryname) %>%
summarise(avg_tier = mean(tier, na.rm=TRUE),
improve_tip = (last(na.omit(tier), default=NA) -
first(na.omit(tier), default=NA)),
change_policy = last(na.omit(p), default=NA) -
first(na.omit(p), default=NA)) %>%
mutate(countryname = countrycode(countryname, "country.name", "country.name"))
# Democracy (Freedom House)
if (!file.exists(file.path(PROJHOME, "data_external", "freedom_house.xls"))) {
fh.url <- paste0("https://freedomhouse.org/sites/default/files/",
"Individual%20Country%20Ratings%20and%20Status%2C%20",
"1973-2015%20%28FINAL%29.xls")
fh.tmp <- file.path(PROJHOME, "data_external", "freedom_house.xls")
download.file(fh.url, fh.tmp)
}
fh.raw <- read_excel(file.path(PROJHOME, "data_external", "freedom_house.xls"),
skip=6)
## DEFINEDNAME: 20 00 00 01 1a 00 00 00 01 00 00 00 00 00 00 07 29 17 00 3b 00 00 00 00 ff ff 00 00 00 00 3b 00 00 00 00 06 00 00 00 ff 00 10
## DEFINEDNAME: 20 00 00 01 1a 00 00 00 01 00 00 00 00 00 00 07 29 17 00 3b 00 00 00 00 ff ff 00 00 00 00 3b 00 00 00 00 06 00 00 00 ff 00 10
## DEFINEDNAME: 20 00 00 01 1a 00 00 00 01 00 00 00 00 00 00 07 29 17 00 3b 00 00 00 00 ff ff 00 00 00 00 3b 00 00 00 00 06 00 00 00 ff 00 10
## DEFINEDNAME: 20 00 00 01 1a 00 00 00 01 00 00 00 00 00 00 07 29 17 00 3b 00 00 00 00 ff ff 00 00 00 00 3b 00 00 00 00 06 00 00 00 ff 00 10
# Calculate the number of years covered in the data (each year has three columns)
num.years <- (ncol(fh.raw) - 1)/3
# Create combinations of all the variables and years
var.years <- expand.grid(var = c('PR', 'CL', 'Status'),
year = 1972:(1972 + num.years - 1))
colnames(fh.raw) <- c('country', paste(var.years$var, var.years$year, sep="_"))
# Split columns and convert to long
fh <- fh.raw %>%
gather(var.year, value, -country) %>%
separate(var.year, into=c("indicator", "year"), sep="_") %>%
filter(!is.na(country)) %>%
spread(indicator, value) %>%
mutate(year = as.numeric(year),
CL = suppressWarnings(as.integer(CL)),
PR = suppressWarnings(as.integer(PR)),
Status = factor(Status, levels=c("NF", "PF", "F"),
labels=c("Not free", "Partially free", "Free"),
ordered=TRUE),
total.freedom = CL + PR,
country.clean = countrycode(country, "country.name", "country.name")) %>%
filter(!is.na(CL) & !is.na(PR)) %>%
# All the cases we're interested in are after 2000, so we can remove these
# problematic double countries
filter(!(country %in% c("Germany, E.", "Germany, W.", "USSR", "Vietnam, N.",
"Vietnam, S.", "Yemen, N.", "Yemen, S."))) %>%
# Again, because we only care about post-2000 Serbia, merge with Yugoslavia
mutate(country.clean = ifelse(country.clean == "Yugoslavia",
"Serbia", country.clean)) %>%
select(-country, country=country.clean)
fh.summary <- fh %>%
filter(year >= 2000) %>%
group_by(country) %>%
summarize(total.freedom = mean(total.freedom, na.rm=TRUE))
# Funding
funding.raw <- read_csv(file.path(PROJHOME, "data_raw", "funding_clean.csv")) %>%
mutate(cowcode = ifelse(country == "Serbia", 555, cowcode),
countryname = countrycode(cowcode, "cown", "country.name"),
countryname = ifelse(cowcode == 555, "Serbia", countryname)) %>%
filter(!is.na(countryname))
funding.all <- funding.raw %>%
group_by(countryname) %>%
summarise(total.funding = sum(amount, na.rm=TRUE),
avg.funding = mean(amount, na.rm=TRUE))
funding.ngos <- funding.raw %>%
filter(recipient_type %in% c("NGO", "NPO")) %>%
group_by(countryname) %>%
summarise(total.funding.ngos = sum(amount, na.rm=TRUE),
avg.funding.ngos = mean(amount, na.rm=TRUE))
responses.full <- responses.all %>%
filter(work.only.us != "Yes") %>%
mutate(work.country.clean = countrycode(work.country,
"country.name", "country.name"),
work.country.clean = ifelse(is.na(work.country),
"Global", work.country.clean),
work.country = work.country.clean) %>%
left_join(tip.change, by=c("work.country" = "countryname")) %>%
left_join(funding.all, by=c("work.country" = "countryname")) %>%
left_join(funding.ngos, by=c("work.country" = "countryname")) %>%
left_join(fh.summary, by=c("work.country" = "country")) %>%
left_join(positivity, by = "Q3.25") %>%
left_join(importance, by = "Q3.19") %>%
left_join(improvement, by = "Q3.26") %>%
mutate(received.funding = ifelse(Q3.18_3 != 1 | is.na(Q3.18_3), FALSE, TRUE),
us.involvement = ifelse(Q3.18_5 != 1 | is.na(Q3.18_5), TRUE, FALSE),
us.hq = ifelse(home.country == "United States", TRUE, FALSE),
Q3.19 = factor(Q3.19, levels=c("Most important actor",
"Somewhat important actor",
"Not an important actor",
"Don't know"),
ordered=TRUE),
Q3.25_collapsed = ifelse(Q3.25 == "Negative", NA, Q3.25)) %>%
mutate(home.region = countrycode(home.country, "country.name", "continent"),
home.region = ifelse(home.country == "Kosovo", "Europe", home.region),
home.region = ifelse(home.country == "TWN", "Asia", home.region),
home.region = ifelse(home.region == "Oceania", "Asia", home.region),
home.region = ifelse(home.region == "Asia", "Asia and Oceania", home.region),
work.region = countrycode(work.country, "country.name", "continent"),
work.region = ifelse(work.country == "Kosovo", "Europe", work.region),
work.region = ifelse(work.country == "TWN", "Asia", work.region),
work.region = ifelse(work.region == "Oceania", "Asia", work.region),
work.region = ifelse(work.region == "Asia", "Asia and Oceania", work.region)) %>%
mutate(home.iso3 = countrycode(home.country, "country.name", "iso3c"),
home.iso3 = ifelse(home.country == "Kosovo", "KOS", home.iso3)) %>%
mutate(work.iso3 = countrycode(work.country, "country.name", "iso3c"),
work.iso3 = ifelse(work.country == "Kosovo", "KOS", work.iso3))
country.indexes <- responses.full %>%
filter(!is.na(work.country)) %>%
group_by(work.country) %>%
# Needs mutate + mutate_each + select + unique because you can't mix
# summarise + summarise_each. See http://stackoverflow.com/a/31815540/120898
mutate(num.responses = n()) %>%
mutate_each(funs(score = mean(., na.rm=TRUE), stdev = sd(., na.rm=TRUE),
n = sum(!is.na(.))),
c(positivity, importance, improvement)) %>%
select(work.country, num.responses, matches("positivity_|importance_|improvement_")) %>%
unique %>%
ungroup() %>%
arrange(desc(num.responses))
# Load map data
if (!file.exists(file.path(PROJHOME, "data_external", "map_data",
"ne_110m_admin_0_countries.VERSION.txt"))) {
map.url <- paste0("http://www.naturalearthdata.com/",
"http//www.naturalearthdata.com/download/110m/cultural/",
"ne_110m_admin_0_countries.zip")
map.tmp <- file.path(PROJHOME, "data_external", basename(map.url))
download.file(map.url, map.tmp)
unzip(map.tmp, exdir=file.path(PROJHOME, "data_external", "map_data"))
unlink(map.tmp)
}
countries.map <- readOGR(file.path(PROJHOME, "data_external", "map_data"),
"ne_110m_admin_0_countries")
## OGR data source with driver: ESRI Shapefile
## Source: "/Users/andrew/Research/••Projects/Human trafficking/Anti-TIP NGOs and the US/data_external/map_data", layer: "ne_110m_admin_0_countries"
## with 177 features
## It has 63 fields
countries.robinson <- spTransform(countries.map, CRS("+proj=robin"))
countries.ggmap <- fortify(countries.robinson, region="iso_a3") %>%
filter(!(id %in% c("ATA", -99))) %>% # Get rid of Antarctica and NAs
mutate(id = ifelse(id == "GRL", "DNK", id)) # Greenland is part of Denmark
# All possible countries (to fix the South Sudan issue)
possible.countries <- data_frame(id = unique(as.character(countries.ggmap$id)))
# Save data
# TODO: Make responses.full more anonymous before making it public
# TODO: Save as CSV and Stata too instead of just R
saveRDS(responses.full, file.path(PROJHOME, "data", "responses_full.rds"))
saveRDS(country.indexes, file.path(PROJHOME, "data", "country_indexes.rds"))
# Useful functions
theme_clean <- function(base_size=9, base_family="Source Sans Pro Light") {
ret <- theme_bw(base_size, base_family) +
theme(panel.background = element_rect(fill="#ffffff", colour=NA),
axis.title.x=element_text(vjust=-0.2), axis.title.y=element_text(vjust=1.5),
title=element_text(vjust=1.2, family="Source Sans Pro Semibold"),
panel.border = element_blank(), axis.line=element_blank(),
panel.grid.minor=element_blank(),
panel.grid.major.y = element_blank(),
panel.grid.major.x = element_line(size=0.25, colour="grey90"),
axis.ticks=element_blank(),
legend.position="bottom",
axis.title=element_text(size=rel(0.8), family="Source Sans Pro Semibold"),
strip.text=element_text(size=rel(0.9), family="Source Sans Pro Semibold"),
strip.background=element_rect(fill="#ffffff", colour=NA),
panel.margin=unit(1, "lines"), legend.key.size=unit(.7, "line"),
legend.key=element_blank())
ret
}
# For maps
theme_blank_map <- function(base_size=12, base_family="Source Sans Pro Light") {
ret <- theme_bw(base_size, base_family) +
theme(panel.background = element_rect(fill="#ffffff", colour=NA),
title=element_text(vjust=1.2, family="Source Sans Pro Semibold"),
panel.border=element_blank(), axis.line=element_blank(),
panel.grid=element_blank(), axis.ticks=element_blank(),
axis.title=element_blank(), axis.text=element_blank(),
legend.text=element_text(size=rel(0.7), family="Source Sans Pro Light"),
legend.title=element_text(size=rel(0.7), family="Source Sans Pro Semibold"),
strip.text=element_text(size=rel(1), family="Source Sans Pro Semibold"))
ret
}
# Return a data frame of counts and proportions for multiple responses
separate.answers.summary <- function(df, cols, labels, total=FALSE) {
cols.to.select <- which(colnames(df) %in% cols)
denominator <- df %>%
select(cols.to.select) %>%
mutate(num.answered = rowSums(., na.rm=TRUE)) %>%
filter(num.answered > 0) %>%
nrow()
df <- df %>%
select(survey.id, cols.to.select) %>%
gather(question, value, -survey.id) %>%
mutate(question = factor(question, labels=labels, ordered=TRUE)) %>%
group_by(question) %>%
summarize(response = sum(value, na.rm=TRUE),
pct = round(response / denominator * 100, 2),
plot.pct = response / denominator)
colnames(df) <- c("Answer", "Responses", "%", "plot.pct")
if (total) {
df <- df %>% select(1:3)
df <- rbind(as.matrix(df), c("Total responses", denominator, "—"))
}
return(list(df=df, denominator=denominator))
}
# Create a character vector of significance stars
add.stars <- function(x) {
as.character(symnum(x, corr = FALSE,
cutpoints = c(0, .001,.01,.05, .1, 1),
symbols = c("***","**","*","."," ")))
}
Calculating the response rate is a little tricky because of all the different ways we sent out invitations for and conducted the survey (LinkedIn, e-mail, phone).
entries.in.list <- 1421 # Organizations in the master NGO list
no.details <- 98 # Organizations that we never tried to make any contact with ever
bounced.invitations <- 132 # Contact information was dead; never saw the survey
not.ngos <- 69 # Organizations that weren't NGOs
duplicates <- 19 # Duplicate entries
orgs.saw.survey <- entries.in.list - no.details - bounced.invitations
viable.entries <- orgs.saw.survey - not.ngos - duplicates
We assume that there were 1191 organizations that saw the link to the survey at least three times. Of those, 69 entries were clearly not NGOs or do not work on human trafficking issues (some were government offices, others only do fundraising, etc.), and most of these organizations responded to the e-mail explaining their situation. 19 entires were duplicates of other entries (generally one entry used the foreign name and one used the English translation).
Given all this, the denominator for our survey’s response rate is 1103.
(num.total.responses <- nrow(responses.full %>% group_by(survey.id) %>% slice(1)))
## [1] 480
num.total.responses / viable.entries
## [1] 0.4351768
How many times did respondents loop through the country questions?
(num.country.orgs <- nrow(responses.full))
## [1] 761
responses.full %>%
group_by(survey.id) %>%
summarise(loops = n()) %>%
do(data.frame(table(.$loops, dnn="Countries"))) %>%
mutate(prop = Freq / sum(Freq))
## Countries Freq prop
## 1 1 415 0.86458333
## 2 4 52 0.10833333
## 3 9 10 0.02083333
## 4 16 3 0.00625000
How did respondents take the survey?
responses.full %>%
group_by(survey.id) %>%
slice(1) %>% # Select each organization's first country
group_by(survey.method) %>%
summarise(num = n()) %>%
mutate(prop = num / sum(num))
## Source: local data frame [3 x 3]
##
## survey.method num prop
## (chr) (int) (dbl)
## 1 LinkedIn 3 0.00625000
## 2 Online 463 0.96458333
## 3 Phone 14 0.02916667
How many respondents answered at least one free response question?
responses.full %>%
select(Q3.10:Q3.17, Q3.24.Text, Q3.30, Q4.1) %T>%
{cat("Number of free response questions:", ncol(.), "\n")} %>%
rowwise() %>% do(wrote.something = !all(is.na(.))) %>%
ungroup() %>% mutate(wrote.something = unlist(wrote.something)) %>%
bind_cols(select(responses.full, survey.id)) %>%
group_by(survey.id) %>%
summarise(wrote.something = ifelse(sum(wrote.something) > 0, TRUE, FALSE)) %T>%
{cat("Number responses:", nrow(.), "\n")} %>%
do(as.data.frame(table(.$wrote.something, dnn="Wrote something"))) %>%
mutate(Percent = Freq / sum(Freq))
## Number of free response questions: 12
## Number responses: 480
## Wrote.something Freq Percent
## 1 FALSE 70 0.1458333
## 2 TRUE 410 0.8541667
Export free response questions for manual analysis
responses.full %>%
select(survey.id, clean.id, home.country, work.country,
Q3.10:Q3.13, Q3.14:Q3.17, Q3.24.Text, Q3.30, Q4.1) %>%
write_csv(path=file.path(PROJHOME, "data", "free_responses.csv"))
How many organizations did not list the US as an anti-TIP actor but later indicated US support?
no.mention.us <- responses.countries %>%
select(survey.id, Q3.6_c2) %>% filter(Q3.6_c2 == 0)# filter(!is.na(Q3.6_c2))
inndicated.us.activity <- responses.full %>%
select(survey.id, Q3.8) %>% filter(Q3.8 == "Yes")
sum(no.mention.us$survey.id %in% inndicated.us.activity$survey.id)
## [1] 42
How many organizations completed the survey after it turned to US-related questions? That’s hard to tell because the US questions were in the loop, and Qualtrics does not keep completion statistics for questions potenitally hidden by display logic.
final.qualtrics.count <- 511
survey.duration.completed <- responses.full %>%
select(survey.id, start.time, end.time) %>%
group_by(survey.id) %>% slice(1) %>% ungroup() %>%
mutate(time.spent = end.time - start.time,
mins.spent = time.spent / 60)
median(as.numeric(survey.duration.completed$mins.spent))
## [1] 21.5
survey.duration <- read_csv(file.path(PROJHOME, "data_raw", "response_times.csv")) %>%
mutate(pct = num / sum(num),
cum.num = cumsum(num),
validish.num = sum(num) - cum.num) %T>%
{print(head(.))}
## Source: local data frame [6 x 5]
##
## minutes_spent num pct cum.num validish.num
## (int) (int) (dbl) (int) (int)
## 1 1 249 0.22656961 249 850
## 2 2 83 0.07552320 332 767
## 3 3 41 0.03730664 373 726
## 4 4 19 0.01728844 392 707
## 5 5 20 0.01819836 412 687
## 6 6 27 0.02456779 439 660
longer.than.five <- filter(survey.duration, minutes_spent==6)$validish.num
We attempted to estimate the post-US completion rate in two ways. First, we counted the number of respondents that spent more than 5 minutes taking the survey (assuming that is sufficient time to make it to the US-focused questions), yielding 660 respondents, which is 149 more than the final number of organizations (511). (Sidenote: 511 is bigger than the final official count of 480 because we filtered out US-only responses completed prior to adding internal validation logic and we removed duplicate or invalid responses prior to analysis) However, because the survey filtered out organizations that only work in the US, most respondents were let out of the survey early. Assuming those who took more than 5 minutes completed the survey, we had a completion rate of 77.4%.
pct.completed <- read_csv(file.path(PROJHOME, "data_raw", "pct_completed.csv")) %>%
mutate(pct = num / sum(num))
more.than.20 <- sum(filter(pct.completed, pct_complete >= 0.2)$num)
Second, we looked at overall completion rates, which again are not entirely accurate because of Qualtrics’ internal question looping logic—i.e., organizations that completed the full survey are only internally marked as completing 20-30% of it. 604 organizations completed at least 20% of the survey. Assuming at least 20% represents approximate survey completion, we had a completion rate of 84.6%.
Thus, given that we filtered out all US-only NGOs in our response rate calculations, and given that we cannot precisely know the true number of survey dropouts, we upwardly bias our completion rate estimate to 90ish%.
Where are these NGOs based?
hq.countries <- responses.full %>%
group_by(survey.id) %>% slice(1) %>% ungroup() %>%
rename(id = home.iso3) %>%
group_by(id) %>%
summarize(num.ngos = n()) %>%
ungroup() %>%
right_join(possible.countries, by="id") %>%
mutate(num.ceiling = ifelse(num.ngos >= 10, 10, num.ngos),
prop = num.ngos / sum(num.ngos, na.rm=TRUE)) %>%
arrange(desc(num.ngos)) %T>%
{print(head(.))}
## Source: local data frame [6 x 4]
##
## id num.ngos num.ceiling prop
## (chr) (int) (dbl) (dbl)
## 1 USA 56 10 0.11940299
## 2 IND 27 10 0.05756930
## 3 NGA 25 10 0.05330490
## 4 GBR 21 10 0.04477612
## 5 CAN 14 10 0.02985075
## 6 THA 10 10 0.02132196
hq.regions <- responses.full %>%
group_by(survey.id) %>% slice(1) %>% ungroup() %>%
filter(!is.na(home.region)) %>%
group_by(home.region) %>%
summarise(num = n()) %>% ungroup() %>%
mutate(prop = num / sum(num)) %T>%
{print(head(.))}
## Source: local data frame [4 x 3]
##
## home.region num prop
## (chr) (int) (dbl)
## 1 Africa 80 0.1670146
## 2 Americas 110 0.2296451
## 3 Asia and Oceania 148 0.3089770
## 4 Europe 141 0.2943633
hq.map <- ggplot(hq.countries, aes(fill=num.ceiling, map_id=id)) +
geom_map(map=countries.ggmap, size=0.15, colour="black") +
expand_limits(x=countries.ggmap$long, y=countries.ggmap$lat) +
coord_equal() +
scale_fill_gradient(low="grey95", high="grey20", breaks=seq(2, 10, 2),
labels=c(paste(seq(2, 8, 2), " "), "10+"),
na.value="#FFFFFF", name="NGOs based in country",
guide=guide_colourbar(ticks=FALSE, barwidth=6)) +
theme_blank_map() +
theme(legend.position="bottom", legend.key.size=unit(0.65, "lines"),
strip.background=element_rect(colour="#FFFFFF", fill="#FFFFFF"))
Where do these NGOs work?
work.countries <- responses.full %>%
rename(id = work.iso3) %>%
group_by(id) %>%
summarize(num.ngos = n()) %>%
ungroup() %>%
right_join(possible.countries, by="id") %>%
mutate(num.ceiling = ifelse(num.ngos >= 10, 10, num.ngos),
prop = num.ngos / sum(num.ngos, na.rm=TRUE)) %>%
arrange(desc(num.ngos)) %T>%
{print(head(.))}
## Source: local data frame [6 x 4]
##
## id num.ngos num.ceiling prop
## (chr) (int) (dbl) (dbl)
## 1 IND 44 10 0.05913978
## 2 NGA 43 10 0.05779570
## 3 GBR 27 10 0.03629032
## 4 THA 24 10 0.03225806
## 5 NPL 22 10 0.02956989
## 6 KHM 20 10 0.02688172
work.regions <- responses.full %>%
filter(!is.na(work.region)) %>%
group_by(work.region) %>%
summarise(num = n()) %>% ungroup() %>%
mutate(prop = num / sum(num)) %T>%
{print(head(.))}
## Source: local data frame [4 x 3]
##
## work.region num prop
## (chr) (int) (dbl)
## 1 Africa 156 0.2058047
## 2 Americas 96 0.1266491
## 3 Asia and Oceania 303 0.3997361
## 4 Europe 203 0.2678100
work.map <- ggplot(work.countries, aes(fill=num.ceiling, map_id=id)) +
geom_map(map=countries.ggmap, size=0.15, colour="black") +
expand_limits(x=countries.ggmap$long, y=countries.ggmap$lat) +
coord_equal() +
scale_fill_gradient(low="grey95", high="grey20", breaks=seq(2, 10, 2),
labels=c(paste(seq(2, 8, 2), " "), "10+"),
na.value="#FFFFFF", name="NGOs working in country",
guide=guide_colourbar(ticks=FALSE, barwidth=6)) +
theme_blank_map() +
theme(legend.position="bottom", legend.key.size=unit(0.65, "lines"),
strip.background=element_rect(colour="#FFFFFF", fill="#FFFFFF"))
Combined maps
fig.maps <- arrangeGrob(hq.map, work.map, nrow=1)
grid.draw(fig.maps)
ggsave(fig.maps, filename=file.path(PROJHOME, "figures", "fig_maps.pdf"),
width=6, height=3, units="in", device=cairo_pdf, scale=1.5)
ggsave(fig.maps, filename=file.path(PROJHOME, "figures", "fig_maps.png"),
width=6, height=3, units="in", type="cairo", scale=1.5)
Side-by-side graph of home vs. work regions
plot.hq <- hq.regions %>%
arrange(num) %>%
mutate(region = factor(home.region, levels=home.region, ordered=TRUE),
prop.nice = sprintf("%.1f%%", prop * 100))
plot.work <- work.regions %>%
mutate(region = factor(work.region, levels=levels(plot.hq$region), ordered=TRUE),
prop.nice = sprintf("%.1f%%", prop * 100))
fig.hq <- ggplot(plot.hq, aes(x=region, y=num)) +
geom_bar(stat="identity") +
geom_text(aes(label = prop.nice), size=2.5, hjust=1.3,
family="Source Sans Pro Light") +
labs(x=NULL, y="NGOs based in region") +
scale_y_continuous(trans="reverse", expand = c(.1, .1)) +
coord_flip(ylim=c(0, 200)) +
theme_clean() +
theme(axis.text.y = element_blank(),
axis.line.y = element_blank(),
plot.margin = unit(c(1,0.5,1,1), "lines"))
fig.work <- ggplot(plot.work, aes(x=region, y=num)) +
geom_bar(stat="identity") +
geom_text(aes(label = prop.nice), size=2.5, hjust=-0.3,
family="Source Sans Pro Light") +
labs(x=NULL, y="NGOs working in region") +
scale_y_continuous(expand = c(.15, .15)) +
coord_flip() +
theme_clean() +
theme(axis.text.y = element_text(hjust=0.5),
axis.line.y = element_blank(),
plot.margin = unit(c(1,1,1,0), "lines"))
fig.locations <- arrangeGrob(fig.hq, fig.work, nrow=1)
grid.draw(fig.locations)
ggsave(fig.locations, filename=file.path(PROJHOME, "figures", "fig_locations.pdf"),
width=5, height=1.5, units="in", device=cairo_pdf, scale=2.5)
ggsave(fig.locations, filename=file.path(PROJHOME, "figures", "fig_locations.png"),
width=5, height=1.5, units="in", type="cairo", scale=2.5)
Where do different regional NGOs work?
responses.full %>%
filter(!is.na(work.region)) %>%
group_by(home.region, work.region) %>%
summarise(num = n()) %>%
group_by(home.region) %>%
mutate(prop.home.region = num / sum(num)) %>%
print
## Source: local data frame [14 x 4]
## Groups: home.region [4]
##
## home.region work.region num prop.home.region
## (chr) (chr) (int) (dbl)
## 1 Africa Africa 108 0.955752212
## 2 Africa Asia and Oceania 4 0.035398230
## 3 Africa Europe 1 0.008849558
## 4 Americas Africa 28 0.135922330
## 5 Americas Americas 92 0.446601942
## 6 Americas Asia and Oceania 69 0.334951456
## 7 Americas Europe 17 0.082524272
## 8 Asia and Oceania Africa 3 0.015000000
## 9 Asia and Oceania Asia and Oceania 195 0.975000000
## 10 Asia and Oceania Europe 2 0.010000000
## 11 Europe Africa 17 0.071129707
## 12 Europe Americas 4 0.016736402
## 13 Europe Asia and Oceania 35 0.146443515
## 14 Europe Europe 183 0.765690377
orgs.only <- responses.full %>%
group_by(survey.id) %>% slice(1) %>% ungroup()
How much time and resources do these NGOs spend on trafficking?
fig.time <- ggplot(data=orgs.only,
aes(x=as.numeric(Q2.1)/100, y=(..count.. / sum(..count..)))) +
geom_histogram(binwidth=0.1) +
labs(x="Proportion of time spent on trafficking", y="Proportion of responses") +
scale_x_continuous(labels=percent, limits=c(0, 1), breaks=seq(0, 1, 0.2)) +
scale_y_continuous(labels=percent, breaks=seq(0, 0.12, 0.02)) +
coord_cartesian(ylim=c(0, 0.125)) +
theme_clean() + theme(panel.grid.major.y = element_line(size=0.25, colour="grey90"))
fig.time
ggsave(fig.time, filename=file.path(PROJHOME, "figures", "fig_time.pdf"),
width=5, height=2, units="in", device=cairo_pdf)
ggsave(fig.time, filename=file.path(PROJHOME, "figures", "fig_time.png"),
width=5, height=2, units="in", type="cairo")
Summary stats of time spent
orgs.only %>%
summarise(avg.time = mean(Q2.1, na.rm=TRUE),
med.time = median(Q2.1, na.rm=TRUE),
min.time = min(Q2.1, na.rm=TRUE),
max.time = max(Q2.1, na.rm=TRUE))
## Source: local data frame [1 x 4]
##
## avg.time med.time min.time max.time
## (dbl) (dbl) (int) (int)
## 1 56.96476 56 1 100
How much do organizations know about trafficking in the countries they work in?
responses.full %>%
group_by(Q3.3) %>%
summarise(num = n()) %>%
filter(!is.na(Q3.3)) %>%
mutate(prop = num / sum(num)) %T>%
{print(sum(.$num))}
## [1] 753
## Source: local data frame [6 x 3]
##
## Q3.3 num prop
## (fctr) (int) (dbl)
## 1 None 8 0.010624170
## 2 Very little 28 0.037184595
## 3 Little 23 0.030544489
## 4 Some 147 0.195219124
## 5 A lot 541 0.718459495
## 6 Don't know 6 0.007968127
What trafficking issues is the organization involved with?
cols <- c("Q2.2_1", "Q2.2_2", "Q2.2_3", "Q2.2_4")
labels <- c("Organ trafficking", "Sex trafficking",
"Labor trafficking", "Other")
issues <- separate.answers.summary(orgs.only, cols, labels)
issues$denominator # Number of responses
## [1] 479
issues$df
## Source: local data frame [4 x 4]
##
## Answer Responses % plot.pct
## (fctr) (int) (dbl) (dbl)
## 1 Organ trafficking 30 6.26 0.06263048
## 2 Sex trafficking 408 85.18 0.85177453
## 3 Labor trafficking 294 61.38 0.61377871
## 4 Other 116 24.22 0.24217119
plot.data <- issues$df %>%
mutate(Answer=factor(Answer, levels=rev(labels), ordered=TRUE))
fig.issues <- ggplot(plot.data, aes(x=Answer, y=Responses)) +
geom_bar(aes(y=plot.pct), stat="identity") +
labs(x=NULL, y=NULL) +
scale_y_continuous(labels = percent,
breaks = seq(0, max(round(plot.data$plot.pct, 1)), by=0.1)) +
coord_flip() + theme_clean()
fig.issues
How many NGOs deal with both sex and labor trafficking?
orgs.only %>%
mutate(sex = ifelse(is.na(Q2.2_2), 0, 1),
labor = ifelse(is.na(Q2.2_3), 0, 1),
both = sex == 1 & labor == 1) %>%
summarise(sex = sum(sex), labor = sum(labor),
num.both = sum(both), prop.both = num.both / issues$denominator)
## Source: local data frame [1 x 4]
##
## sex labor num.both prop.both
## (dbl) (dbl) (int) (dbl)
## 1 408 294 243 0.5073069
Other responses about issues NGOs deal with
# orgs.only %>% filter(!is.na(Q2.2_4_TEXT)) %>%
# select(clean.id, Q2.2_4_TEXT) %>% View
Which kinds of victims do NGOs help?
cols <- c("Q2.3_1", "Q2.3_2", "Q2.3_3")
labels <- c("Children", "Adults", "Other")
victims <- separate.answers.summary(orgs.only, cols, labels)
victims$denominator # Number of responses
## [1] 478
victims$df
## Source: local data frame [3 x 4]
##
## Answer Responses % plot.pct
## (fctr) (int) (dbl) (dbl)
## 1 Children 335 70.08 0.7008368
## 2 Adults 318 66.53 0.6652720
## 3 Other 75 15.69 0.1569038
plot.data <- victims$df %>%
mutate(Answer=factor(Answer, levels=rev(labels), ordered=TRUE))
fig.victims <- ggplot(plot.data, aes(x=Answer, y=Responses)) +
geom_bar(aes(y=plot.pct), stat="identity") +
labs(x=NULL, y=NULL) +
scale_y_continuous(labels = percent,
breaks = seq(0, max(round(plot.data$plot.pct, 1)), by=0.1)) +
coord_flip() + theme_clean()
fig.victims
Other responses about victims NGOs deal with
# orgs.only %>% filter(!is.na(Q2.3_3_TEXT)) %>%
# select(clean.id, Q2.3_3_TEXT) %>% View
How are the types of victims distributed between the main trafficking issues?
victims.issues <- orgs.only %>%
select(organs=Q2.2_1, sex=Q2.2_2, labor=Q2.2_3, other.issue=Q2.2_4,
children=Q2.3_1, adults=Q2.3_2, other.victims=Q2.3_3) %>%
mutate_each(funs(!is.na(.)))
victims.issues %>%
filter(children) %>%
summarise(num = n(),
sex = sum(sex), sex.prop = sex / n(),
labor = sum(labor), labor.prop = labor / n())
## Source: local data frame [1 x 5]
##
## num sex sex.prop labor labor.prop
## (int) (int) (dbl) (int) (dbl)
## 1 335 291 0.8686567 211 0.6298507
victims.issues %>%
filter(adults) %>%
summarise(num = n(),
sex = sum(sex), sex.prop = sex / n(),
labor = sum(labor), labor.prop = labor / n())
## Source: local data frame [1 x 5]
##
## num sex sex.prop labor labor.prop
## (int) (int) (dbl) (int) (dbl)
## 1 318 284 0.8930818 212 0.6666667
Which efforts do NGOs focus on?
cols <- c("Q2.4_1", "Q2.4_2", "Q2.4_3", "Q2.4_4", "Q2.4_5")
labels <- c("Prevention and education", "Prosecutions and legal issues",
"Victim protection", "Victim assistance", "Other")
efforts <- separate.answers.summary(orgs.only, cols, labels)
efforts$denominator # Number of responses
## [1] 479
efforts$df
## Source: local data frame [5 x 4]
##
## Answer Responses % plot.pct
## (fctr) (int) (dbl) (dbl)
## 1 Prevention and education 398 83.09 0.8308977
## 2 Prosecutions and legal issues 188 39.25 0.3924843
## 3 Victim protection 248 51.77 0.5177453
## 4 Victim assistance 340 70.98 0.7098121
## 5 Other 128 26.72 0.2672234
plot.data <- efforts$df %>%
arrange(plot.pct) %>%
mutate(Answer=factor(Answer, levels=Answer, ordered=TRUE))
fig.efforts <- ggplot(plot.data, aes(x=Answer, y=Responses)) +
geom_bar(aes(y=plot.pct), stat="identity") +
labs(x=NULL, y=NULL) +
scale_y_continuous(labels = percent,
breaks = seq(0, max(round(plot.data$plot.pct, 1)), by=0.1)) +
coord_flip() + theme_clean()
fig.efforts
Other responses about efforts NGOs engage in
# orgs.only %>% filter(!is.na(Q2.4_5_TEXT)) %>%
# select(clean.id, Q2.4_5_TEXT) %>% View
How are the types of efforts distributed between the main trafficking issues?
efforts.issues <- orgs.only %>%
select(prevention=Q2.4_1, legal=Q2.4_2, protection=Q2.4_3, assistance=Q2.4_4,
sex=Q2.2_2, labor=Q2.2_3) %>%
mutate_each(funs(!is.na(.)))
efforts.issues %>%
filter(prevention) %>%
summarise(num = n(),
sex = sum(sex), sex.prop = sex / n(),
labor = sum(labor), labor.prop = labor / n())
## Source: local data frame [1 x 5]
##
## num sex sex.prop labor labor.prop
## (int) (int) (dbl) (int) (dbl)
## 1 398 338 0.8492462 259 0.6507538
efforts.issues %>%
filter(legal) %>%
summarise(num = n(),
sex = sum(sex), sex.prop = sex / n(),
labor = sum(labor), labor.prop = labor / n())
## Source: local data frame [1 x 5]
##
## num sex sex.prop labor labor.prop
## (int) (int) (dbl) (int) (dbl)
## 1 188 168 0.893617 125 0.6648936
efforts.issues %>%
filter(protection) %>%
summarise(num = n(),
sex = sum(sex), sex.prop = sex / n(),
labor = sum(labor), labor.prop = labor / n())
## Source: local data frame [1 x 5]
##
## num sex sex.prop labor labor.prop
## (int) (int) (dbl) (int) (dbl)
## 1 248 219 0.8830645 172 0.6935484
efforts.issues %>%
filter(assistance) %>%
summarise(num = n(),
sex = sum(sex), sex.prop = sex / n(),
labor = sum(labor), labor.prop = labor / n())
## Source: local data frame [1 x 5]
##
## num sex sex.prop labor labor.prop
## (int) (int) (dbl) (int) (dbl)
## 1 340 297 0.8735294 218 0.6411765
Which institutions have been active in anti-TIP work?
cols <- c("Q3.5_1", "Q3.5_2", "Q3.5_3", "Q3.5_4", "Q3.5_5")
labels <- c("The national government", "NGOs and civil society",
"Foreign governments", "International organizations", "Other")
other.actors <- separate.answers.summary(responses.full, cols, labels)
other.actors$denominator # Number of responses
## [1] 747
other.actors$df
## Source: local data frame [5 x 4]
##
## Answer Responses % plot.pct
## (fctr) (int) (dbl) (dbl)
## 1 The national government 477 63.86 0.6385542
## 2 NGOs and civil society 709 94.91 0.9491299
## 3 Foreign governments 301 40.29 0.4029451
## 4 International organizations 487 65.19 0.6519411
## 5 Other 111 14.86 0.1485944
plot.data <- other.actors$df %>%
arrange(plot.pct) %>%
mutate(Answer=factor(Answer, levels=Answer, ordered=TRUE))
fig.other.actors <- ggplot(plot.data, aes(x=Answer, y=Responses)) +
geom_bar(aes(y=plot.pct), stat="identity") +
labs(x=NULL, y=NULL) +
scale_y_continuous(labels = percent,
breaks = seq(0, max(round(plot.data$plot.pct, 1)), by=0.1)) +
coord_flip() + theme_clean()
fig.other.actors
Other responses
others <- responses.full %>% select(survey.id, starts_with("Q3.5_5_")) %>%
filter(!is.na(Q3.5_5_TEXT)) %>%
# Convert values to numeric
mutate_each(funs(as.numeric(levels(.))[.]), -c(survey.id, Q3.5_5_TEXT))
cols <- c(c("Q3.5_5_LawEnforcement", "Q3.5_5_Education",
"Q3.5_5_DomesticGovernment", "Q3.5_5_ReligiousGroups",
"Q3.5_5_Embassies", "Q3.5_5_InternationalGroups",
"Q3.5_5_NGOs", "Q3.5_5_Other", "Q3.5_5_Media",
"Q3.5_5_ExperienceofSurviviors",
"Q3.5_5_PrivateSector", "Q3.5_5_Unions"))
labels <- c("Law Enforcement", "Education",
"Domestic Government", "Religious Groups",
"Embassies", "International Groups",
"NGOs", "Other", "Media",
"Experience of Surviviors",
"Private Sector", "Unions")
(others.summary <- separate.answers.summary(others, cols, labels))
## $df
## Source: local data frame [12 x 4]
##
## Answer Responses % plot.pct
## (fctr) (dbl) (dbl) (dbl)
## 1 Law Enforcement 10 10.99 0.10989011
## 2 Education 4 4.40 0.04395604
## 3 Domestic Government 30 32.97 0.32967033
## 4 Religious Groups 17 18.68 0.18681319
## 5 Embassies 2 2.20 0.02197802
## 6 International Groups 12 13.19 0.13186813
## 7 NGOs 13 14.29 0.14285714
## 8 Other 10 10.99 0.10989011
## 9 Media 2 2.20 0.02197802
## 10 Experience of Surviviors 3 3.30 0.03296703
## 11 Private Sector 2 2.20 0.02197802
## 12 Unions 2 2.20 0.02197802
##
## $denominator
## [1] 91
How hard is the government working?
responses.full %>%
group_by(Q3.20) %>%
summarise(num = n()) %>%
filter(!is.na(Q3.20)) %>%
mutate(prop = num / sum(num))
## Source: local data frame [6 x 3]
##
## Q3.20 num prop
## (fctr) (int) (dbl)
## 1 Not hard at all 77 0.1036339
## 2 Not too hard 183 0.2462988
## 3 Somewhat hard 310 0.4172275
## 4 Very hard 112 0.1507402
## 5 Extremely hard 21 0.0282638
## 6 Don't know 40 0.0538358
Have NGOs used the TIP report to talk with any of these groups?
cols <- c("Q3.21_1", "Q3.21_2", "Q3.21_3", "Q3.21_4")
labels <- c("National government", "Another government",
"Other NGOs", "Other")
tip.discuss <- separate.answers.summary(responses.full, cols, labels)
tip.discuss$denominator # Number of responses
## [1] 543
tip.discuss$df
## Source: local data frame [4 x 4]
##
## Answer Responses % plot.pct
## (fctr) (int) (dbl) (dbl)
## 1 National government 281 51.75 0.5174954
## 2 Another government 106 19.52 0.1952118
## 3 Other NGOs 430 79.19 0.7918969
## 4 Other 107 19.71 0.1970534
plot.data <- tip.discuss$df %>%
arrange(plot.pct) %>%
mutate(Answer=factor(Answer, levels=Answer, ordered=TRUE))
fig.tip.discuss <- ggplot(plot.data, aes(x=Answer, y=Responses)) +
geom_bar(aes(y=plot.pct), stat="identity") +
labs(x=NULL, y=NULL) +
scale_y_continuous(labels = percent,
breaks = seq(0, max(round(plot.data$plot.pct, 1)), by=0.1)) +
coord_flip() + theme_clean()
fig.tip.discuss
Do members of the government or ruling party sit on the NGO’s board?
responses.full %>%
group_by(Q3.27) %>%
summarise(num = n()) %>%
filter(!is.na(Q3.27)) %>%
mutate(prop = num / sum(num))
## Source: local data frame [3 x 3]
##
## Q3.27 num prop
## (fctr) (int) (dbl)
## 1 No 638 0.86567164
## 2 Yes 50 0.06784261
## 3 Don't know 49 0.06648575
For those that said yes, is the NGO required to have a government official sit on the board?
responses.full %>%
group_by(Q3.28) %>%
summarise(num = n()) %>%
filter(!is.na(Q3.28)) %>%
mutate(prop = num / sum(num))
## Source: local data frame [3 x 3]
##
## Q3.28 num prop
## (fctr) (int) (dbl)
## 1 No 30 0.61224490
## 2 Yes 17 0.34693878
## 3 Don't know 2 0.04081633
How restricted does the NGO feel by the host government?
responses.full %>%
group_by(Q3.29) %>%
summarise(num = n()) %>%
filter(!is.na(Q3.29)) %>%
mutate(prop = num / sum(num)) %T>%
{cat("Number of responses:", sum(.$num), "\n")}
## Number of responses: 734
## Source: local data frame [6 x 3]
##
## Q3.29 num prop
## (fctr) (int) (dbl)
## 1 Not restricted 270 0.36784741
## 2 Very little restricted 155 0.21117166
## 3 A little restricted 63 0.08583106
## 4 Somewhat restricted 125 0.17029973
## 5 Very restricted 62 0.08446866
## 6 Don't know 59 0.08038147
Which countries do NGOs say are restrictive?
responses.full %>%
filter(Q3.29 %in% c("Somewhat restricted", "Very restricted")) %>%
group_by(work.country) %>%
summarise(num = n()) %>%
arrange(desc(num)) %T>%
{cat("Number of countries:", nrow(.), "\n")}
## Number of countries: 66
## Source: local data frame [66 x 2]
##
## work.country num
## (chr) (int)
## 1 India 12
## 2 Nigeria 10
## 3 Thailand 10
## 4 Cambodia 9
## 5 Congo, the Democratic Republic of the 8
## 6 Malaysia 7
## 7 Nepal 6
## 8 Singapore 6
## 9 Colombia 5
## 10 Iraq 5
## .. ... ...
Explanations of restrictions
# responses.full %>% filter(!is.na(Q3.30)) %>%
# select(clean.id, Q3.30) %>% View
# Select just the columns that have cowcodes embedded in them
active.embassies.raw <- responses.countries %>%
select(contains("_c", ignore.case=FALSE)) %>%
mutate_each(funs(as.numeric(levels(.))[.])) # Convert values to numeric
# Select only the rows where they responded (i.e. not all columns are NA)
num.responses <- active.embassies.raw %>%
rowwise() %>% do(all.missing = all(!is.na(.))) %>%
ungroup() %>% mutate(all.missing = unlist(all.missing)) %>%
summarise(total = sum(all.missing))
# Tidy cowcode columns and summarize most commonly mentioned countries
active.embassies <- active.embassies.raw %>%
gather(country.raw, num) %>%
group_by(country.raw) %>% summarise(num = sum(num, na.rm=TRUE)) %>%
mutate(country.raw = str_replace(country.raw, "Q.*c", ""),
country = countrycode(country.raw, "cown", "country.name"),
country = ifelse(country.raw == "2070", "European Union", country)) %>%
ungroup() %>% mutate(prop = num / num.responses$total,
prop.nice = sprintf("%.1f%%", prop * 100))
Which embassies or foreign governments NGOs were reported as active partners in the fight against human trafficking?
active.embassies.top <- active.embassies %>%
arrange(num) %>% select(-country.raw) %>%
filter(num > 10) %>%
mutate(country = factor(country, levels=country, ordered=TRUE)) %>%
arrange(desc(num))
active.embassies %>% arrange(desc(num)) %>%
select(-country.raw) %>% filter(num > 10)
## Source: local data frame [13 x 4]
##
## num country prop prop.nice
## (dbl) (chr) (dbl) (chr)
## 1 335 United States 0.76834862 76.8%
## 2 58 United Kingdom 0.13302752 13.3%
## 3 37 Netherlands 0.08486239 8.5%
## 4 37 France 0.08486239 8.5%
## 5 35 Norway 0.08027523 8.0%
## 6 33 Sweden 0.07568807 7.6%
## 7 33 European Union 0.07568807 7.6%
## 8 29 Switzerland 0.06651376 6.7%
## 9 27 Australia 0.06192661 6.2%
## 10 25 Germany 0.05733945 5.7%
## 11 25 Italy 0.05733945 5.7%
## 12 18 Canada 0.04128440 4.1%
## 13 11 Philippines 0.02522936 2.5%
nrow(active.embassies) # Number of countries mentioned
## [1] 64
num.responses$total # Total responses
## [1] 436
# Most active embassies
# Save Q3.7 to a CSV for hand coding
most.active <- responses.countries %>%
select(Q3.7) %>%
filter(!is.na(Q3.7))
write_csv(most.active, path=file.path(PROJHOME, "data",
"most_active_WILL_BE_OVERWRITTEN.csv"))
# Read in hand-coded CSV
if (file.exists(file.path(PROJHOME, "data", "most_active.csv"))) {
most.active <- read_csv(file.path(PROJHOME, "data", "most_active.csv"))
} else {
stop("data/most_active.csv is missing")
}
# Split comma-separated countries, unnest them into multiple rows, and
# summarize most active countries
most.active.clean <- most.active %>%
transform(clean = strsplit(clean, ",")) %>%
unnest(clean) %>%
mutate(clean = str_trim(clean)) %>%
group_by(clean) %>%
summarise(total = n()) %>%
mutate(prop = total / nrow(most.active),
prop.nice = sprintf("%.1f%%", prop * 100))
Which countries or embassies have been the most active?
most.active.clean %>% arrange(desc(total))
## Source: local data frame [40 x 4]
##
## clean total prop prop.nice
## (chr) (int) (dbl) (chr)
## 1 United States 188 0.70149254 70.1%
## 2 None 16 0.05970149 6.0%
## 3 European Union 14 0.05223881 5.2%
## 4 All 12 0.04477612 4.5%
## 5 Switzerland 8 0.02985075 3.0%
## 6 Australia 7 0.02611940 2.6%
## 7 Italy 7 0.02611940 2.6%
## 8 United Kingdom 7 0.02611940 2.6%
## 9 Netherlands 6 0.02238806 2.2%
## 10 Norway 6 0.02238806 2.2%
## .. ... ... ... ...
nrow(most.active.clean) - 1 # Subtract one because of "None"s
## [1] 39
Over the last 10–15 years, has the United States or its embassy been active in the fight against human trafficking in X?
responses.countries$Q3.8 %>% table %>% print %>% prop.table
## .
## No Yes Don't know
## 49 434 219
## .
## No Yes Don't know
## 0.06980057 0.61823362 0.31196581
Side-by-side graph of active countries + most active countries
plot.data <- active.embassies.top %>%
bind_rows(data_frame(num=0, country=c("All", "None"),
prop=0, prop.nice="")) %>%
arrange(num) %>%
mutate(country = factor(country, levels=country, ordered=TRUE))
plot.data.active <- most.active.clean %>%
filter(clean %in% plot.data$country) %>%
mutate(country = factor(clean, levels=levels(plot.data$country), ordered=TRUE))
fig.active <- ggplot(plot.data, aes(x=country, y=num)) +
geom_bar(stat="identity") +
geom_text(aes(label = prop.nice), size=3.5, hjust=1.3,
family="Source Sans Pro Light") +
labs(x=NULL, y="Number of times country was mentioned as a partner in anti-TIP work") +
scale_y_continuous(breaks=seq(0, max(active.embassies$num), by=25),
trans="reverse", expand = c(.1, .1)) +
coord_flip() +
theme_clean() +
theme(axis.text.y = element_blank(),
axis.line.y = element_blank(),
plot.margin = unit(c(1,0.5,1,1), "lines"))
fig.most.active <- ggplot(plot.data.active, aes(x=country, y=total)) +
geom_bar(stat="identity") +
geom_text(aes(label = prop.nice), size=3.5, hjust=-0.3,
family="Source Sans Pro Light") +
labs(x=NULL, y="Number of times country was mentioned as the most active partner in anti-TIP work") +
scale_y_continuous(expand = c(.15, .15)) +
coord_flip() +
theme_clean() +
theme(axis.text.y = element_text(hjust=0.5),
axis.line.y = element_blank(),
plot.margin = unit(c(1,1,1,0), "lines"))
fig.embassies <- arrangeGrob(fig.active, fig.most.active, nrow=1)
grid.draw(fig.embassies)
ggsave(fig.embassies, filename=file.path(PROJHOME, "figures", "fig_embassies.pdf"),
width=5, height=2, units="in", device=cairo_pdf, scale=2.5)
ggsave(fig.embassies, filename=file.path(PROJHOME, "figures", "fig_embassies.png"),
width=5, height=2, units="in", type="cairo", scale=2.5)
saveRDS(active.embassies, file.path(PROJHOME, "data", "active_embassies.rds"))
saveRDS(most.active.clean, file.path(PROJHOME, "data", "most_active_embassies.rds"))
Actual US activities
cols <- c("Q3.9_1", "Q3.9_2", "Q3.9_3", "Q3.9_4", "Q3.9_5",
"Q3.9_6", "Q3.9_7", "Q3.9_8", "Q3.9_9", "Q3.9_10")
labels <- c("Asking for legislation", "Convening conferences or workshops",
"Raising awareness", "Providing resources or funding",
"Increasing government attention", "Training government officials",
"Contributing to a government action plan", "Other", "Don't know",
"The US has not been involved in trafficking issues")
activities <- separate.answers.summary(responses.countries, cols, labels)
activities$denominator # Number of responses
## [1] 702
activities$df
## Source: local data frame [10 x 4]
##
## Answer Responses % plot.pct
## (fctr) (int) (dbl) (dbl)
## 1 Asking for legislation 204 29.06 0.29059829
## 2 Convening conferences or workshops 263 37.46 0.37464387
## 3 Raising awareness 255 36.32 0.36324786
## 4 Providing resources or funding 269 38.32 0.38319088
## 5 Increasing government attention 267 38.03 0.38034188
## 6 Training government officials 183 26.07 0.26068376
## 7 Contributing to a government action plan 141 20.09 0.20085470
## 8 Other 56 7.98 0.07977208
## 9 Don't know 29 4.13 0.04131054
## 10 The US has not been involved in trafficking issues 246 35.04 0.35042735
plot.data <- activities$df %>%
mutate(Answer=factor(Answer, levels=rev(labels), ordered=TRUE))
fig.activities <- ggplot(plot.data, aes(x=Answer, y=Responses)) +
geom_bar(aes(y=plot.pct), stat="identity") +
labs(x=NULL, y=NULL) +
scale_y_continuous(labels = percent,
breaks = seq(0, max(round(plot.data$plot.pct, 1)), by=0.1)) +
coord_flip() + theme_clean()
fig.activities
ggsave(fig.activities, filename=file.path(PROJHOME, "figures", "fig_activities.pdf"),
width=6.5, height=5, units="in", device=cairo_pdf)
ggsave(fig.activities, filename=file.path(PROJHOME, "figures", "fig_activities.png"),
width=6.5, height=5, units="in", type="cairo")
plot.data <- responses.full %>%
group_by(Q3.19) %>%
summarize(num = n()) %>%
na.omit() %>%
mutate(prop = num / sum(num),
prop.nice = sprintf("%.1f%%", prop * 100),
Q3.19 = factor(Q3.19, levels=rev(levels(Q3.19)), ordered=TRUE))
plot.data
## Source: local data frame [4 x 4]
##
## Q3.19 num prop prop.nice
## (fctr) (int) (dbl) (chr)
## 1 Most important actor 181 0.2681481 26.8%
## 2 Somewhat important actor 224 0.3318519 33.2%
## 3 Not an important actor 84 0.1244444 12.4%
## 4 Don't know 186 0.2755556 27.6%
fig.us_importance <- ggplot(plot.data, aes(x=Q3.19, y=prop)) +
geom_bar(stat="identity") +
labs(x=NULL, y=NULL) +
scale_y_continuous(labels = percent,
breaks = seq(0, max(round(plot.data$num, 1)), by=0.1)) +
coord_flip() + theme_clean()
fig.us_importance
Average importance by country
importance.plot <- country.indexes %>%
filter(num.responses >= 10) %>%
arrange(importance_score) %>%
mutate(country_label = factor(work.country, levels=unique(work.country),
labels=paste0(work.country, " (", num.responses, ")"),
ordered=TRUE))
fig.avg_importance <- ggplot(importance.plot, aes(x=country_label, y=importance_score)) +
geom_pointrange(aes(ymax=importance_score + importance_stdev,
ymin=importance_score - importance_stdev)) +
labs(x="Country (number of responses)",
y="Importance of the US in anti-TIP efforts (mean)") +
scale_y_discrete(breaks=c(0, 1, 2), labels=c("Not important", "Somewhat important", "Most important")) +
coord_flip(ylim=c(0, 2)) +
theme_clean() + theme(legend.position="bottom")
fig.avg_importance
ggsave(fig.avg_importance, filename=file.path(PROJHOME, "figures", "fig_avg_importance.pdf"),
width=6.5, height=3, units="in", device=cairo_pdf)
ggsave(fig.avg_importance, filename=file.path(PROJHOME, "figures", "fig_avg_importance.png"),
width=6.5, height=3, units="in", type="cairo")
df.importance <- responses.full %>%
select(Q3.19, work.country, change_policy, avg_tier, improve_tip, change_policy,
importance, received.funding, us.involvement, total.funding,
total.freedom, us.hq, time.spent=Q2.1) %>%
filter(!is.na(Q3.19)) %>%
filter(Q3.19 != "Don't know") %>%
mutate(importance_factor = factor(Q3.19, ordered=FALSE),
log.total.funding = log1p(total.funding),
time.spent = as.numeric(time.spent))
Average tier doesn’t show much because it doesn’t show any changes in time—just how bad the country is in general?
importance.means <- df.importance %>%
group_by(Q3.19) %>%
summarize(avg_points = mean(avg_tier, na.rm=TRUE),
var_points = var(avg_tier, na.rm=TRUE)) %>%
print
## Source: local data frame [3 x 3]
##
## Q3.19 avg_points var_points
## (fctr) (dbl) (dbl)
## 1 Most important actor 2.079981 0.1145449
## 2 Somewhat important actor 1.919886 0.2370952
## 3 Not an important actor 1.833792 0.3500821
Plot group means and distributions
fig.importance <- ggplot(df.importance, aes(x=Q3.19, y=avg_tier)) +
geom_violin(fill="grey90") +
geom_point(alpha=0.05, show.legend=FALSE) +
geom_point(data=importance.means, aes(x=Q3.19, y=avg_points), size=5, show.legend=FALSE) +
labs(x="Opinion of US importance", y="Average TIP tier rating") +
coord_flip() + theme_clean()
fig.importance
Those means appear slightly different from each other. Is that really the case? Check with ANOVA, which assumes homogenous variance across groups. Throw every possible test at it—if null is rejected (p < 0.05 or whatever) then variance is likely heterogenous: (helpful reference)
bartlett.test(avg_tier ~ importance_factor, data=df.importance)
##
## Bartlett test of homogeneity of variances
##
## data: avg_tier by importance_factor
## Bartlett's K-squared = 41.793, df = 2, p-value = 8.411e-10
car::leveneTest(avg_tier ~ importance_factor, data=df.importance)
## Levene's Test for Homogeneity of Variance (center = median)
## Df F value Pr(>F)
## group 2 16.768 9.085e-08 ***
## 485
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
fligner.test(avg_tier ~ importance_factor, data=df.importance) # Uses median
##
## Fligner-Killeen test of homogeneity of variances
##
## data: avg_tier by importance_factor
## Fligner-Killeen:med chi-squared = 25.726, df = 2, p-value = 2.592e-06
kruskal.test(avg_tier ~ importance_factor, data=df.importance) # Nonparametric
##
## Kruskal-Wallis rank sum test
##
## data: avg_tier by importance_factor
## Kruskal-Wallis chi-squared = 14.476, df = 2, p-value = 0.0007189
All of those p-values are tiny, so it’s clear that variance is not the same across groups. However, there’s a rule of thumb (super detailed example) that ANOVA is robust to heterogeneity of variance as long as the largest variance is less than four times the smallest variance.
Given that rule of thumb, the variance here isn’t that much of an issue
df.importance %>% group_by(importance_factor) %>%
summarise(variance = var(avg_tier, na.rm=TRUE)) %>%
do(data_frame(ratio = max(.$variance) / min(.$variance)))
## Source: local data frame [1 x 1]
##
## ratio
## (dbl)
## 1 3.056287
It would be cool to use Bayesian ANOVA to account for non-homogenous variances (see John Kruschke’s evangelizing), since it handles violations of ANOVA assumptions nicely. However, in his example, the ratio of min/max variance is huge, so it does lead to big differences in results:
#
# read_csv("http://www.indiana.edu/~kruschke/DoingBayesianDataAnalysis/Programs/NonhomogVarData.csv") %>%
# group_by(Group) %>%
# summarise(variance = var(Y)) %>%
# do(data_frame(ratio = max(.$variance) / min(.$variance)))
# # ratio = 64
#
With the variance issue handled, run the ANOVA:
importance.aov <- aov(avg_tier ~ importance_factor, data=df.importance)
summary(importance.aov)
## Df Sum Sq Mean Sq F value Pr(>F)
## importance_factor 2 4.29 2.1464 10.18 4.69e-05 ***
## Residuals 485 102.31 0.2109
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 1 observation deleted due to missingness
There is some significant difference between groups. Look at pairwise comparisons between all the groups to (kind of) decompose that finding:
(importance.pairs <- TukeyHSD(importance.aov, "importance_factor"))
## Tukey multiple comparisons of means
## 95% family-wise confidence level
##
## Fit: aov(formula = avg_tier ~ importance_factor, data = df.importance)
##
## $importance_factor
## diff lwr upr p adj
## Somewhat important actor-Most important actor -0.16009574 -0.2681214 -0.05207010 0.0015592
## Not an important actor-Most important actor -0.24618961 -0.3887411 -0.10363808 0.0001688
## Not an important actor-Somewhat important actor -0.08609388 -0.2243247 0.05213698 0.3091384
Plot the differences:
importance.pairs.plot <- data.frame(importance.pairs$importance_factor) %>%
mutate(pair = row.names(.),
pair = factor(pair, levels=pair, ordered=TRUE),
stars = add.stars(p.adj))
fig.importance.pairs <- ggplot(importance.pairs.plot,
aes(x=pair, y=diff, ymax=upr, ymin=lwr)) +
geom_hline(yintercept=0) +
geom_text(aes(label=stars), nudge_x=0.25) +
geom_pointrange() +
theme_clean() + coord_flip()
fig.importance.pairs
Another way of checking group means in non-homogenous data is to use ordinal logistic regression. Here’s an ordered logit and corresponding predicted probabilities:
model.importance <- ordinal::clm(Q3.19 ~ avg_tier, data=df.importance,
link="logit", Hess=TRUE)
summary(model.importance)
## formula: Q3.19 ~ avg_tier
## data: df.importance
##
## link threshold nobs logLik AIC niter max.grad cond.H
## logit flexible 488 -492.06 990.13 5(0) 1.35e-12 1.9e+02
##
## Coefficients:
## Estimate Std. Error z value Pr(>|z|)
## avg_tier -0.8202 0.1860 -4.41 1.03e-05 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Threshold coefficients:
## Estimate Std. Error z value
## Most important actor|Somewhat important actor -2.15388 0.38218 -5.636
## Somewhat important actor|Not an important actor 0.01755 0.36592 0.048
## (1 observation deleted due to missingness)
# Predicted probabilities
newdata <- data_frame(avg_tier = seq(0, 3, 0.1))
pred.importance <- predict(model.importance, newdata, interval=TRUE)
# Create plot data
pred.plot.lower <- cbind(newdata, pred.importance$lwr) %>%
gather(importance, lwr, -c(1:ncol(newdata)))
pred.plot.upper <- cbind(newdata, pred.importance$upr) %>%
gather(importance, upr, -c(1:ncol(newdata)))
pred.plot.data <- cbind(newdata, pred.importance$fit) %>%
gather(importance, importance_prob, -c(1:ncol(newdata))) %>%
left_join(pred.plot.lower, by=c("avg_tier", "importance")) %>%
left_join(pred.plot.upper, by=c("avg_tier", "importance"))
importance.colors <- c("grey20", "grey40", "grey60", "grey80")
ggplot(pred.plot.data, aes(x=avg_tier, y=importance_prob)) +
geom_ribbon(aes(ymax=upr, ymin=lwr, fill=importance),
alpha=0.2) +
geom_line(aes(colour=importance), size=2) +
scale_y_continuous(labels=percent) +
labs(x="Average tier rating in country",
y="Predicted probability of assigning importance") +
# scale_fill_manual(values=importance.colors, name=NULL) +
# scale_colour_manual(values=importance.colors, name=NULL) +
theme_clean()
Opinions of importance are not related to changes in TIP score. The average Improvement in TIP rating is the same for each possible answer of importance.
ggplot(df.importance, aes(x=Q3.19, y=improve_tip)) +
geom_violin(fill="grey90") +
geom_point(stat="summary", fun.y="mean", size=5) +
labs(x="Opinion of US", y="Improvement in TIP tier rating") +
coord_flip() + theme_clean()
Variance is equal in all groups:
kruskal.test(improve_tip ~ importance_factor, data=df.importance)
##
## Kruskal-Wallis rank sum test
##
## data: improve_tip by importance_factor
## Kruskal-Wallis chi-squared = 0.46164, df = 2, p-value = 0.7939
ANOVA shows no differences:
change.anova <- aov(improve_tip ~ importance_factor, data=df.importance)
summary(change.anova)
## Df Sum Sq Mean Sq F value Pr(>F)
## importance_factor 2 0.04 0.02066 0.066 0.936
## Residuals 485 151.18 0.31171
## 1 observation deleted due to missingness
TukeyHSD(change.anova)
## Tukey multiple comparisons of means
## 95% family-wise confidence level
##
## Fit: aov(formula = improve_tip ~ importance_factor, data = df.importance)
##
## $importance_factor
## diff lwr upr p adj
## Somewhat important actor-Most important actor -0.017342616 -0.1486578 0.1139725 0.9482552
## Not an important actor-Most important actor 0.003157064 -0.1701275 0.1764416 0.9989891
## Not an important actor-Somewhat important actor 0.020499680 -0.1475327 0.1885321 0.9556725
Opinions of importance vary slightly by changes in Cho policy scores. Respondents who indicated that the US was more important tended to work in countries with greater changes in TIP policy.
ggplot(df.importance, aes(x=Q3.19, y=change_policy)) +
geom_violin(fill="grey90") +
geom_point(stat="summary", fun.y="mean", size=5) +
labs(x="Opinion of US", y="Change in TIP policy index") +
coord_flip() + theme_clean()
Variance is equal in all groups:
kruskal.test(change_policy ~ importance_factor, data=df.importance)
##
## Kruskal-Wallis rank sum test
##
## data: change_policy by importance_factor
## Kruskal-Wallis chi-squared = 2.6473, df = 2, p-value = 0.2662
ANOVA shows no differences
cho.change.anova <- aov(change_policy ~ importance_factor, data=df.importance)
summary(cho.change.anova) # (⌐○Ϟ○) ♥ \(•◡•)/
## Df Sum Sq Mean Sq F value Pr(>F)
## importance_factor 2 23 11.629 1.708 0.182
## Residuals 485 3301 6.807
## 1 observation deleted due to missingness
TukeyHSD(cho.change.anova)
## Tukey multiple comparisons of means
## 95% family-wise confidence level
##
## Fit: aov(formula = change_policy ~ importance_factor, data = df.importance)
##
## $importance_factor
## diff lwr upr p adj
## Somewhat important actor-Most important actor -0.2940564 -0.9077067 0.3195938 0.4981544
## Not an important actor-Most important actor -0.6208235 -1.4306014 0.1889545 0.1698655
## Not an important actor-Somewhat important actor -0.3267670 -1.1120010 0.4584669 0.5909347
Organizations that have been received funding from the US are more likely to consider the US to play an important role in the countries they work in.
funding.table <- df.importance %>%
xtabs(~ importance_factor + received.funding, .) %>% print
## received.funding
## importance_factor FALSE TRUE
## Most important actor 92 89
## Somewhat important actor 180 44
## Not an important actor 80 4
There’s an overall significant difference (though two of the cells are really small here)
(funding.chi <- chisq.test(funding.table))
##
## Pearson's Chi-squared test
##
## data: funding.table
## X-squared = 70.478, df = 2, p-value = 4.966e-16
# Cramer's V for standardized measure of association
assocstats(funding.table)
## X^2 df P(> X^2)
## Likelihood Ratio 75.083 2 0.0000e+00
## Pearson 70.478 2 4.4409e-16
##
## Phi-Coefficient : NA
## Contingency Coeff.: 0.355
## Cramer's V : 0.38
# Components of chi-squared
(components <- funding.chi$residuals^2)
## received.funding
## importance_factor FALSE TRUE
## Most important actor 11.252970 28.912741
## Somewhat important actor 2.181868 5.605968
## Not an important actor 6.310414 16.213617
round(1-pchisq(components, funding.chi$parameter), 3)
## received.funding
## importance_factor FALSE TRUE
## Most important actor 0.004 0.000
## Somewhat important actor 0.336 0.061
## Not an important actor 0.043 0.000
# Visualize differences
mosaic(funding.table,
labeling_args=list(set_varnames=c(received.funding="Received US funding",
importance_factor="Opinion of US"),
gp_labels=(gpar(fontsize=8))),
gp_varnames=gpar(fontsize=10, fontface=2))
Opinions of importance are strongly associated with US TIP funding given to a country. Organizations are more likely to think the US is an important actor if they work in countries receiving more anti-TIP funding.
ggplot(df.importance, aes(x=Q3.19, y=log.total.funding)) +
geom_violin(fill="grey90") +
geom_point(alpha=0.05) +
geom_point(stat="summary", fun.y="mean", size=5) +
labs(x="Opinion of US", y="Total TIP funding to country (logged)") +
scale_y_continuous(labels=trans_format("exp", dollar_format())) +
coord_flip() + theme_clean()
Variance is not equal in all groups:
kruskal.test(log.total.funding ~ importance_factor, data=df.importance)
##
## Kruskal-Wallis rank sum test
##
## data: log.total.funding by importance_factor
## Kruskal-Wallis chi-squared = 13.209, df = 2, p-value = 0.001355
Ratio is 3ish, which is below 4, so heterogenous variance is okayish:
df.importance %>% group_by(importance_factor) %>%
summarise(variance = var(log.total.funding, na.rm=TRUE)) %>%
do(data_frame(ratio = max(.$variance) / min(.$variance)))
## Source: local data frame [1 x 1]
##
## ratio
## (dbl)
## 1 3.072576
ANOVA shows significant differences:
funding.anova <- aov(log.total.funding ~ importance_factor, data=df.importance)
summary(funding.anova)
## Df Sum Sq Mean Sq F value Pr(>F)
## importance_factor 2 792 396.2 13.43 2.12e-06 ***
## Residuals 475 14010 29.5
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 11 observations deleted due to missingness
(funding.pairs <- TukeyHSD(funding.anova))
## Tukey multiple comparisons of means
## 95% family-wise confidence level
##
## Fit: aov(formula = log.total.funding ~ importance_factor, data = df.importance)
##
## $importance_factor
## diff lwr upr p adj
## Somewhat important actor-Most important actor -1.439429 -2.734085 -0.1447724 0.0249897
## Not an important actor-Most important actor -3.721300 -5.416179 -2.0264210 0.0000011
## Not an important actor-Somewhat important actor -2.281871 -3.920600 -0.6431425 0.0032599
See those differences
funding.pairs.plot <- data.frame(funding.pairs$importance_factor) %>%
mutate(pair = row.names(.),
pair = factor(pair, levels=pair, ordered=TRUE),
stars = add.stars(p.adj))
fig.funding.pairs <- ggplot(funding.pairs.plot,
aes(x=pair, y=diff, ymax=upr, ymin=lwr)) +
geom_hline(yintercept=0) +
geom_text(aes(label=stars), nudge_x=0.25) +
geom_pointrange() +
theme_clean() + coord_flip()
fig.funding.pairs
US importance appears to be associated with the level of democracy in a country. NGOs working in countries with worse democracy (higher numbers of the total freedom scale) are more likely to see the US as the most important anti-TIP actor in that country. Or, rather, on average total freedom is worse in countries where NGOs indicate the US as the most important actor.
fig.importance.freedom <- ggplot(df.importance, aes(x=Q3.19, y=total.freedom)) +
# geom_violin(fill="grey90") +
geom_point(alpha=0.05, size=0.25) +
geom_point(stat="summary", fun.y="mean", size=2) +
scale_y_continuous(breaks=seq(2, 14, by=4)) +
labs(x=NULL,
y="Total freedom (political rights + civil liberties; higher is worse)") +
coord_flip() + theme_clean(6)
fig.importance.freedom
Variance is not equal in all groups:
kruskal.test(total.freedom ~ importance_factor, data=df.importance)
##
## Kruskal-Wallis rank sum test
##
## data: total.freedom by importance_factor
## Kruskal-Wallis chi-squared = 15.364, df = 2, p-value = 0.000461
Ratio between min and max variance is low, so we’re okay:
df.importance %>% group_by(importance_factor) %>%
summarise(variance = var(total.freedom, na.rm=TRUE)) %>%
do(data_frame(ratio = max(.$variance) / min(.$variance)))
## Source: local data frame [1 x 1]
##
## ratio
## (dbl)
## 1 1.568149
ANOVA shows significant differences:
democracy.anova <- aov(total.freedom ~ importance_factor, data=df.importance)
summary(democracy.anova)
## Df Sum Sq Mean Sq F value Pr(>F)
## importance_factor 2 152 76.19 7.145 0.000876 ***
## Residuals 481 5129 10.66
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 5 observations deleted due to missingness
(democracy.pairs <- TukeyHSD(democracy.anova))
## Tukey multiple comparisons of means
## 95% family-wise confidence level
##
## Fit: aov(formula = total.freedom ~ importance_factor, data = df.importance)
##
## $importance_factor
## diff lwr upr p adj
## Somewhat important actor-Most important actor -1.0699242 -1.842759 -0.2970898 0.0034746
## Not an important actor-Most important actor -1.3499606 -2.367121 -0.3328001 0.0054317
## Not an important actor-Somewhat important actor -0.2800365 -1.262853 0.7027804 0.7811117
View the differences:
democracy.pairs.plot <- data.frame(democracy.pairs$importance_factor) %>%
mutate(pair = row.names(.),
pair = factor(pair, levels=pair, ordered=TRUE),
stars = add.stars(p.adj))
fig.democracy.pairs <- ggplot(democracy.pairs.plot,
aes(x=pair, y=diff, ymax=upr, ymin=lwr)) +
geom_hline(yintercept=0) +
geom_text(aes(label=stars), nudge_x=0.25) +
geom_pointrange() +
theme_clean() + coord_flip()
fig.democracy.pairs
The time NGOs spend on trafficking issues does not appear to be associated with their opinion of US importance.
ggplot(df.importance, aes(x=Q3.19, y=time.spent)) +
geom_violin(fill="grey90") +
geom_point(alpha=0.05) +
geom_point(stat="summary", fun.y="mean", size=5) +
labs(x="Opinion of US", y="Time spent on trafficking issues") +
coord_flip() + theme_clean()
Variance is not equal in all groups:
kruskal.test(time.spent ~ importance_factor, data=df.importance)
##
## Kruskal-Wallis rank sum test
##
## data: time.spent by importance_factor
## Kruskal-Wallis chi-squared = 9.1275, df = 2, p-value = 0.01042
Ratio between min and max variance is low, so we’re okay:
df.importance %>% group_by(importance_factor) %>%
summarise(variance = var(time.spent, na.rm=TRUE)) %>%
do(data_frame(ratio = max(.$variance) / min(.$variance)))
## Source: local data frame [1 x 1]
##
## ratio
## (dbl)
## 1 1.320093
ANOVA shows some small overall signifcant differences, but when decomposed, that effect is coming only from the tiny “Don’t know-Somewhat important actor” difference.
time.anova <- aov(time.spent ~ importance_factor, data=df.importance)
summary(time.anova)
## Df Sum Sq Mean Sq F value Pr(>F)
## importance_factor 2 8630 4315 4.288 0.0143 *
## Residuals 460 462854 1006
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 26 observations deleted due to missingness
(time.pairs <- TukeyHSD(time.anova))
## Tukey multiple comparisons of means
## 95% family-wise confidence level
##
## Fit: aov(formula = time.spent ~ importance_factor, data = df.importance)
##
## $importance_factor
## diff lwr upr p adj
## Somewhat important actor-Most important actor 9.315559 1.676616 16.954503 0.0120248
## Not an important actor-Most important actor 7.380250 -2.720553 17.481053 0.1995107
## Not an important actor-Somewhat important actor -1.935309 -11.792461 7.921842 0.8892154
Organizations that have been involved with the US are more likely to consider the US to play an important role in the countries they work in.
involvement.table <- df.importance %>%
xtabs(~ importance_factor + us.involvement, .) %>% print
## us.involvement
## importance_factor FALSE TRUE
## Most important actor 29 152
## Somewhat important actor 62 162
## Not an important actor 49 35
There’s an overall significant difference
(involvement.chi <- chisq.test(involvement.table))
##
## Pearson's Chi-squared test
##
## data: involvement.table
## X-squared = 50.451, df = 2, p-value = 1.109e-11
# Cramer's V for standardized measure of association
assocstats(involvement.table)
## X^2 df P(> X^2)
## Likelihood Ratio 47.961 2 3.8489e-11
## Pearson 50.451 2 1.1086e-11
##
## Phi-Coefficient : NA
## Contingency Coeff.: 0.306
## Cramer's V : 0.321
# Components of chi-squared
(components <- involvement.chi$residuals^2)
## us.involvement
## importance_factor FALSE TRUE
## Most important actor 10.04928320 4.03123108
## Somewhat important actor 0.07080281 0.02840228
## Not an important actor 25.88657975 10.38430133
1-pchisq(components, involvement.chi$parameter)
## us.involvement
## importance_factor FALSE TRUE
## Most important actor 6.573942e-03 1.332384e-01
## Somewhat important actor 9.652179e-01 9.858992e-01
## Not an important actor 2.392217e-06 5.560036e-03
# Visualize differences
mosaic(involvement.table,
labeling_args=list(set_varnames=c(us.involvement="US involvement",
importance_factor="Opinion of US"),
gp_labels=(gpar(fontsize=8))),
gp_varnames=gpar(fontsize=10, fontface=2))
NGOs with headquarters in the US are not significantly different from their foreign counterparts in their opinions of the importance of the US.
hq.table <- df.importance %>%
xtabs(~ importance_factor + us.hq, .) %>% print
## us.hq
## importance_factor FALSE TRUE
## Most important actor 167 14
## Somewhat important actor 208 16
## Not an important actor 72 12
There’s no overall significant difference
(hq.chi <- chisq.test(hq.table))
##
## Pearson's Chi-squared test
##
## data: hq.table
## X-squared = 4.237, df = 2, p-value = 0.1202
# Cramer's V is really low
assocstats(hq.table)
## X^2 df P(> X^2)
## Likelihood Ratio 3.7482 2 0.15349
## Pearson 4.2370 2 0.12021
##
## Phi-Coefficient : NA
## Contingency Coeff.: 0.093
## Cramer's V : 0.093
# Components of chi-squared
(components <- hq.chi$residuals^2)
## us.hq
## importance_factor FALSE TRUE
## Most important actor 0.01444603 0.15374708
## Somewhat important actor 0.05124435 0.54538625
## Not an important actor 0.29821951 3.17390760
1-pchisq(components, hq.chi$parameter)
## us.hq
## importance_factor FALSE TRUE
## Most important actor 0.9928030 0.9260070
## Somewhat important actor 0.9747033 0.7613264
## Not an important actor 0.8614746 0.2045478
# Visualize differences
mosaic(hq.table,
labeling_args=list(set_varnames=c(us.involvement="US involvement",
importance_factor="Opinion of US"),
gp_labels=(gpar(fontsize=8))),
gp_varnames=gpar(fontsize=10, fontface=2))
Important caveat: Respondents were only asked about their opinions of the US’s work (Q3.25) if they indicated that the US was a somewhat important actor or the most important actor (Q3.19)
df.positivity <- responses.full %>%
select(Q3.25=Q3.25_collapsed, work.country, change_policy, avg_tier,
improve_tip, change_policy, importance, received.funding, us.involvement,
total.funding, total.freedom, us.hq, time.spent=Q2.1) %>%
filter(!is.na(Q3.25)) %>%
filter(Q3.25 != "Don't know") %>%
mutate(positivity.factor = factor(Q3.25, ordered=FALSE),
log.total.funding = log1p(total.funding),
time.spent = as.numeric(time.spent))
plot.data <- df.positivity %>%
group_by(Q3.25) %>%
summarize(num = n()) %>%
na.omit() %>%
mutate(prop = num / sum(num),
prop.nice = sprintf("%.1f%%", prop * 100),
Q3.25 = factor(Q3.25, levels=rev(levels(df.positivity$positivity.factor)),
ordered=TRUE))
plot.data
## Source: local data frame [2 x 4]
##
## Q3.25 num prop prop.nice
## (fctr) (int) (dbl) (chr)
## 1 Mixed 87 0.2464589 24.6%
## 2 Positive 266 0.7535411 75.4%
fig.us_positivity <- ggplot(plot.data, aes(x=Q3.25, y=prop)) +
geom_bar(stat="identity") +
labs(x=NULL, y=NULL) +
scale_y_continuous(labels = percent,
breaks = seq(0, max(round(plot.data$num, 1)), by=0.1)) +
coord_flip() + theme_clean()
fig.us_positivity
Average positivity by country
positivity.plot <- country.indexes %>%
filter(num.responses >= 10,
positivity_score > 0) %>%
arrange(positivity_score) %>%
mutate(country_label = factor(work.country, levels=unique(work.country),
labels=paste0(work.country, " (", num.responses, ")"),
ordered=TRUE))
fig.avg_positivity <- ggplot(positivity.plot, aes(x=country_label, y=positivity_score)) +
geom_pointrange(aes(ymax=positivity_score + positivity_stdev,
ymin=positivity_score - positivity_stdev)) +
labs(x="Country (number of responses)",
y="Positivity of the US in anti-TIP efforts (mean)") +
scale_y_discrete(breaks=c(-1, 0, 1), labels=c("Negative", "Mixed", "Positive")) +
coord_flip(ylim=c(-1, 1)) +
theme_clean() + theme(legend.position="bottom")
fig.avg_positivity
ggsave(fig.avg_positivity, filename=file.path(PROJHOME, "figures", "fig_avg_positivity.pdf"),
width=6.5, height=3, units="in", device=cairo_pdf)
ggsave(fig.avg_positivity, filename=file.path(PROJHOME, "figures", "fig_avg_positivity.png"),
width=6.5, height=3, units="in", type="cairo")
ggplot(df.positivity, aes(x=positivity.factor, y=avg_tier)) +
geom_violin(fill="grey90") +
geom_point(alpha=0.05, show.legend=FALSE) +
geom_point(stat="summary", fun.y="mean", size=5) +
labs(x="Opinion of US", y="Average TIP tier rating") +
coord_flip() + theme_clean()
t.test(avg_tier ~ positivity.factor, data=df.positivity)
##
## Welch Two Sample t-test
##
## data: avg_tier by positivity.factor
## t = 2.3949, df = 163.7, p-value = 0.01776
## alternative hypothesis: true difference in means is not equal to 0
## 95 percent confidence interval:
## 0.02011186 0.20908668
## sample estimates:
## mean in group Mixed mean in group Positive
## 2.079365 1.964766
NGOs who have positive opinions of the US are more likely to work in countries where the TIP rating has (slightly) decreased on average between 2000 and 2015. This may be because assigning a worse TIP rating to a country represents increased US diplomatic and economic pressure—it is a possible sign that NGOs like scorecard diplomacy.
ggplot(df.positivity, aes(x=positivity.factor, y=improve_tip)) +
geom_violin(fill="grey90") +
geom_point(stat="summary", fun.y="mean", size=5) +
labs(x="Opinion of US", y="Improvement in TIP tier rating") +
coord_flip() + theme_clean()
Difference is significant
t.test(improve_tip ~ positivity.factor, data=df.positivity)
##
## Welch Two Sample t-test
##
## data: improve_tip by positivity.factor
## t = 5.0027, df = 165.24, p-value = 1.434e-06
## alternative hypothesis: true difference in means is not equal to 0
## 95 percent confidence interval:
## 0.2059118 0.4744178
## sample estimates:
## mean in group Mixed mean in group Positive
## 0.1609195 -0.1792453
In contrast to the changes in actual TIP scores, NGOs that work in countries that show greater improvement in overall TIP policies are more likely to have a positive opinion of the US, perhaps because they are happy about the actual on-the-ground improvements.
ggplot(df.positivity, aes(x=positivity.factor, y=change_policy)) +
geom_violin(fill="grey90") +
geom_point(stat="summary", fun.y="mean", size=5) +
labs(x="Opinion of US", y="Change in TIP policy index") +
coord_flip() + theme_clean()
Difference is significant
t.test(change_policy ~ positivity.factor, data=df.positivity)
##
## Welch Two Sample t-test
##
## data: change_policy by positivity.factor
## t = -4.7936, df = 191.37, p-value = 3.282e-06
## alternative hypothesis: true difference in means is not equal to 0
## 95 percent confidence interval:
## -1.8601699 -0.7756141
## sample estimates:
## mean in group Mixed mean in group Positive
## 1.689655 3.007547
df.positivity %>% group_by(positivity.factor) %>%
summarise(variance = var(change_policy, na.rm=TRUE)) %>%
do(data_frame(ratio = max(.$variance) / min(.$variance)))
## Source: local data frame [1 x 1]
##
## ratio
## (dbl)
## 1 1.730756
funding.table.pos <- df.positivity %>%
xtabs(~ positivity.factor + received.funding, .) %>% print
## received.funding
## positivity.factor FALSE TRUE
## Mixed 60 27
## Positive 178 88
Distribution isn’t significantly different
(funding.chi.pos <- chisq.test(funding.table.pos))
##
## Pearson's Chi-squared test with Yates' continuity correction
##
## data: funding.table.pos
## X-squared = 0.049326, df = 1, p-value = 0.8242
# Cramer's V for standardized measure of association
assocstats(funding.table.pos)
## X^2 df P(> X^2)
## Likelihood Ratio 0.12593 1 0.72269
## Pearson 0.12521 1 0.72345
##
## Phi-Coefficient : 0.019
## Contingency Coeff.: 0.019
## Cramer's V : 0.019
# Components of chi-squared
(components <- funding.chi.pos$residuals^2)
## received.funding
## positivity.factor FALSE TRUE
## Mixed 0.03073872 0.06361578
## Positive 0.01005364 0.02080667
round(1-pchisq(components, funding.chi.pos$parameter), 3)
## received.funding
## positivity.factor FALSE TRUE
## Mixed 0.861 0.801
## Positive 0.920 0.885
# Visualize differences
mosaic(funding.table.pos,
labeling_args=list(set_varnames=c(received.funding="Received US funding",
positivity.factor="Opinion of US"),
gp_labels=(gpar(fontsize=8))),
gp_varnames=gpar(fontsize=10, fontface=2))
ggplot(df.positivity, aes(x=positivity.factor, y=log.total.funding)) +
geom_violin(fill="grey90") +
geom_point(alpha=0.05) +
geom_point(stat="summary", fun.y="mean", size=5) +
labs(x="Opinion of US", y="Total TIP funding to country (logged)") +
scale_y_continuous(labels=trans_format("exp", dollar_format())) +
coord_flip() + theme_clean()
Difference is significant
t.test(log.total.funding ~ positivity.factor, data=df.positivity)
##
## Welch Two Sample t-test
##
## data: log.total.funding by positivity.factor
## t = 2.0355, df = 180.58, p-value = 0.04326
## alternative hypothesis: true difference in means is not equal to 0
## 95 percent confidence interval:
## 0.03269673 2.10229470
## sample estimates:
## mean in group Mixed mean in group Positive
## 14.58364 13.51615
ggplot(df.positivity, aes(x=positivity.factor, y=total.freedom)) +
geom_violin(fill="grey90") +
geom_point(alpha=0.05) +
geom_point(stat="summary", fun.y="mean", size=5) +
labs(x="Opinion of US",
y="Total freedom (political rights + civil liberties; higher is worse)") +
coord_flip() + theme_clean()
Difference is not quite significant
t.test(total.freedom ~ positivity.factor, data=df.positivity)
##
## Welch Two Sample t-test
##
## data: total.freedom by positivity.factor
## t = 2.2627, df = 159.19, p-value = 0.025
## alternative hypothesis: true difference in means is not equal to 0
## 95 percent confidence interval:
## 0.1042515 1.5353129
## sample estimates:
## mean in group Mixed mean in group Positive
## 7.301495 6.481713
The time NGOs spend on trafficking issues does not appear to be associated with their opinion of US importance.
ggplot(df.positivity, aes(x=positivity.factor, y=time.spent)) +
geom_violin(fill="grey90") +
geom_point(alpha=0.05) +
geom_point(stat="summary", fun.y="mean", size=5) +
labs(x="Opinion of US", y="Time spent on trafficking issues") +
coord_flip() + theme_clean()
Difference is no significant:
t.test(time.spent ~ positivity.factor, data=df.positivity)
##
## Welch Two Sample t-test
##
## data: time.spent by positivity.factor
## t = 0.51224, df = 145.22, p-value = 0.6093
## alternative hypothesis: true difference in means is not equal to 0
## 95 percent confidence interval:
## -5.668462 9.634573
## sample estimates:
## mean in group Mixed mean in group Positive
## 60.37349 58.39044
involvement.table.pos <- df.positivity %>%
xtabs(~ positivity.factor + us.involvement, .) %>% print
## us.involvement
## positivity.factor FALSE TRUE
## Mixed 21 66
## Positive 59 207
There’s no significant difference
(involvement.chi.pos <- chisq.test(involvement.table.pos))
##
## Pearson's Chi-squared test with Yates' continuity correction
##
## data: involvement.table.pos
## X-squared = 0.053396, df = 1, p-value = 0.8173
# Tiny Cramer's V
assocstats(involvement.table.pos)
## X^2 df P(> X^2)
## Likelihood Ratio 0.14190 1 0.70639
## Pearson 0.14332 1 0.70500
##
## Phi-Coefficient : 0.02
## Contingency Coeff.: 0.02
## Cramer's V : 0.02
# Components of chi-squared
(components <- involvement.chi.pos$residuals^2)
## us.involvement
## positivity.factor FALSE TRUE
## Mixed 0.083524226 0.024475964
## Positive 0.027318074 0.008005296
1-pchisq(components, involvement.chi.pos$parameter)
## us.involvement
## positivity.factor FALSE TRUE
## Mixed 0.7725771 0.8756799
## Positive 0.8687222 0.9287065
# Visualize differences
mosaic(involvement.table.pos,
labeling_args=list(set_varnames=c(us.involvement="US involvement",
positivity.factor="Opinion of US"),
gp_labels=(gpar(fontsize=8))),
gp_varnames=gpar(fontsize=10, fontface=2))
NGOs with headquarters in the US are not significantly different from their foreign counterparts in their opinions of the US in general.
hq.table.pos <- df.positivity %>%
xtabs(~ positivity.factor + us.hq, .) %>% print
## us.hq
## positivity.factor FALSE TRUE
## Mixed 81 6
## Positive 248 18
There’s no overall significant difference, but some of the cells are too small
(hq.chi.pos <- chisq.test(hq.table.pos))
##
## Pearson's Chi-squared test with Yates' continuity correction
##
## data: hq.table.pos
## X-squared = 3.6639e-31, df = 1, p-value = 1
# Cramer's V is really, really low
assocstats(hq.table.pos)
## X^2 df P(> X^2)
## Likelihood Ratio 0.0017335 1 0.96679
## Pearson 0.0017386 1 0.96674
##
## Phi-Coefficient : 0.002
## Contingency Coeff.: 0.002
## Cramer's V : 0.002
# Components of chi-squared
(components <- hq.chi.pos$residuals^2)
## us.hq
## positivity.factor FALSE TRUE
## Mixed 8.907435e-05 1.221061e-03
## Positive 2.913334e-05 3.993695e-04
1-pchisq(components, hq.chi.pos$parameter)
## us.hq
## positivity.factor FALSE TRUE
## Mixed 0.9924697 0.9721246
## Positive 0.9956934 0.9840560
# Visualize differences
mosaic(hq.table.pos,
labeling_args=list(set_varnames=c(us.involvement="US involvement",
positivity.factor="Opinion of US"),
gp_labels=(gpar(fontsize=8))),
gp_varnames=gpar(fontsize=10, fontface=2))
Does opinion of the US vary by: * Whether an NGO focuses on certain types of work?
In which countries does the US seem to have had more collaboration with NGOs?
CHECK: Opinions are not driven by cooptation - look at chapter 1 for boomerang type stuff - cooptation by donors - so in this case, the NGOs aren’t just being bought out?
Drop “don’t know”s - that leads to a t test for positivity
Create a nice simple summary table (like the one from Mike Ward’s paper) with graphs
# # Type of work
# # TODO: work.country is not the most reliable identifier---there be NAs
# # Q2.2_X
# asdf <- responses.full %>%
# select(survey.id, matches("Q2.2_\\d$")) %>%
# gather(type_of_work, value, -survey.id) %>%
# mutate(type_of_work = factor(type_of_work, levels=paste0("Q2.2_", seq(1:4)),
# labels=c("Organs", "Sex", "Labor", "Other"),
# ordered=TRUE))
# asdf %>%
# group_by(type_of_work) %>%
# summarise(bloop = n(),
# derp = sum(value, na.rm=TRUE),
# asdf = derp / n())
# # Should be 30, 408, 294, 116 with 479 total?
# #' Does NGO experience on the ground match improvements reported by the State Department?
# #' * Find country averages of government improvement, etc. - then show that X
# #' number of countries show improvement, etc.
# #' * Report by organization and by country - how many countries has the US had
# #' a positive influence + how many NGOs say the US has had a positive
# #' influence
# full <- left_join(country.indexes, tip.change,
# by=c("work.country" = "countryname")) %>%
# filter(num.responses >= 10) %>%
# mutate(country_label = ifelse(num.responses >= 10, work.country, ""))
# ggplot(full, aes(x=improvement_score, y=change_policy, label=work.country)) +
# geom_point() + geom_text(vjust=1.5) +
# geom_hline(yintercept=0) +
# scale_x_continuous(limits=c(0, 1)) +
# scale_y_continuous(limits=c(-2, 6))
# ggplot(country.indexes, aes(x=work.country, y=improvement_score)) +
# geom_bar(stat="identity") +
# coord_flip()
# #' Compare improvement scores with actual changes in TIP score to get a sense
# #' of if NGO experiences reflect changes in rankings
# ggplot(country.indexes, aes(x=work.country, y=positivity_score)) +
# geom_bar(stat="identity") +
# coord_flip()
# # Importance opinions
# importance.opinions <- responses.all %>%
# filter(Q3.19 == "Not an important actor") %>%
# select(survey.id, clean.id, Q3.19, contains("TEXT"), Q4.1)
# responses.all$Q3.19 %>%
# table %>% print %>% prop.table
# responses.countries %>%
# xtabs(~ Q3.25 + Q3.26, .)
# ggplot(responses.orgs, aes(x = Q1.5.factor)) + geom_bar() +
# labs(x = Hmisc::label(responses.orgs$Q1.5))
# # Importance of US
# asdf <- responses.all %>%
# select(clean.id, Q1.2, Q3.8, Q3.6, Q3.7)
# inconsistent.no <- c(1020, 1152, 1226, 1267, 1323, 1405, 1515)
# inconsistent.dont.know <- c(1051, 1512)
# qwer <- asdf %>%
# mutate(us.active = ifelse(clean.id %in% c(inconsistent.no, inconsistent.dont.know),
# "Yes", as.character(Q3.8)))
# qwer$us.active %>% table %>% print %>% prop.table
# sdfg <- qwer %>% group_by(clean.id) %>%
# summarize(said.no = ifelse(any(us.active == "No", na.rm=TRUE), TRUE, FALSE))
# sdfg$said.no %>% table %>% print %>% prop.table
# # US importance and positivity
# # Importance of report
# responses.orgs$Q2.5 %>% table %>% print %>% prop.table
# responses.countries$Q3.23 %>% table %>% print %>% prop.table
# heard.of.tip <- responses.countries %>%
# left_join(responses.orgs, by="survey.id") %>%
# filter(Q2.5 == "Yes") %>%
# group_by(survey.id) %>%
# mutate(know.score = ifelse(Q3.22 == "Don't know", FALSE, TRUE)) %>%
# select(know.score) %>% unique
# heard.of.tip$know.score %>% table %>% print %>% prop.table
# # Opinions of report
# opinions <- responses.all %>%
# select(clean.id, Q1.2, home.country, work.country, Q3.21_1, Q3.21_4_TEXT, Q3.24.Text)
# not.used.tip.ids <- c(1094, 1099, 1106, 1114, 1157, 1221, 1244, 1269,
# 1314, 1330, 1354, 1357, 1393, 1425)
# not.used.tip <- responses.all %>%
# mutate(no.response = ifelse(is.na(Q3.21_1) & is.na(Q3.21_2) &
# is.na(Q3.21_3) & is.na(Q3.21_4), TRUE, FALSE),
# explicit.no = ifelse(clean.id %in% not.used.tip.ids, TRUE, FALSE)) %>%
# select(clean.id, Q1.2, Q3.21_1, Q3.21_2, Q3.21_3, Q3.21_4, no.response, explicit.no) %>%
# group_by(clean.id) %>%
# summarize(no.response = ifelse(sum(no.response) > 0, TRUE, FALSE),
# explicit.no = ifelse(sum(explicit.no) > 0, TRUE, FALSE))
# not.used.tip$explicit.no %>% table